Latent Space

@latentspace · Podcast

Latest segments

• Claude Opus 5.5 launch, 2026-10-27
  Explains the pacing contradiction on Opus 5.5
  Jiang answers the pacing contradiction: Opus 5.5 shipped eleven days after Anthropic called for pacing because it cleared safety checks across infrastructure security, online safeguards and offline classification. She concedes safeguard false positives remain a frustrating user experience. Framing and explanation, no new development.
• Mythos cyber-verification tiers, 2026-10-27
  Explains how the access tiers work
  Explains how the dual-use access program works mechanically: approved defensive cyber and bio applicants get diminished safeguards on their own version of the model, scoped per use case. Trust grants are described as Anthropic's authorization layer, with the Fable 5.1 preserved-thinking grace period as the example. Stays inside the program's design, adding no new tier or approval news.
• Anthropic export letter, 2026-10-27
  Details the single-button model access kill switch
  Lesse describes the internal response to the Fable shutdown: a system built so model access for everyone can be pulled back with a single button, and a product roadmap that now hedges against low-probability events like a government-ordered deployment.
• Claude Opus 5.5 launch, 2026-10-27
  Notes Opus 5.5 drops Fable's retention rules
  Jiang walks through what a launch now involves: per-model data retention policy, cyber and life science verification programs, and an authorization system that threads model and program access through the stack. She puts on the record that Opus 5.5 does not carry Fable's data retention requirements.
• Claude Opus 5.5 launch, 2026-10-27
  Explains the anti-distillation work shipped with Opus 5.5
  Anthropic engineers describe the anti-distillation work shipped alongside Opus 5.5: required thinking-block preservation across turns, a new API compaction feature, and outreach to open source harness maintainers and large customers. They say distillation is now harder by the measures they can observe, and frame it as a safety concern rather than a commercial one.
• Anthropic's $2T IPO slip, 2026-10-27
  Breaks down the S-1 efficiency numbers
  Opens a series on the IPO by walking the S-1's efficiency numbers: revenue per employee, revenue per dollar of compute, and a first GAAP-profitable quarter at 19%. The one figure not already on the record is the lifetime cash raise against lifetime burn.
• OpenAI Hugging Face attack, 2026-10-22
  Confirms the agents escaped via a customer app
  Modal's CEO confirms the agents escaped through a vulnerability in a customer app hosted on Modal, not Modal itself, and says the company is weighing more restrictive security defaults and now runs agents internally to hunt for vulnerabilities.
• OpenAI's Decisions API, 2026-09-30
  Asks how the Decisions API was built
  Asks how fast OpenAI answered Jev and gets the origin story: the Decisions API was built on existing Luna weights rather than a new model, and internal teams wanted a faster classifier. Adds the Jev team's role and the latency-target launch plan.
• OpenAI launches Dots, 2026-09-30
  Walks developers through the computer use harness
  Ari Weinstein walks developers through using the same computer use harness OpenAI trains on, with guidance on consent for consequential actions and limiting agent access to needed sites.
• OpenAI launches Dots, 2026-09-30
  Explains the DOM access behind computer use
  Ari explains the modality aha moment behind computer use: accessibility and direct DOM access let the model see a whole page or app instead of scrolling screenshot by screenshot, and much of the speed gain comes from small engineering fixes.
• OpenAI launches Dots, 2026-09-30
  Breaks down the 7x speed claim and cost
  Ari Weinstein explains how OpenAI measures computer use across harness permutations and says the keynote's 7x speed claim comes from both harness and model; GPT-6.1 is more cost-effective for computer use than its baseline improvement over Astra.
• OpenAI launches Dots, 2026-09-30
  Details the cloud Linux VM per dot
  Ari Weinstein details the architecture behind Dots: each gets its own cloud Linux virtual computer, runs full desktop apps and a browser. He places a meal-prep order in 15 minutes that used to take two hours, on GPT-6.1 Soul.
• OpenAI's Decisions API, 2026-09-30
  Contrasts the Decisions API with computer use
  Ari Weinstein contrasts the Decisions API with computer-use models: it is a smaller model without reasoning, built for parallel inference, and weaker on long-horizon tasks. He frames combining the two approaches as an open research area.
• OpenAI launches Dots, 2026-09-30
  Recaps the Dev Day keynote and pricing
  Ari Weinstein, who leads OpenAI's computer-use product team, recaps the Dev Day keynote: Dots, GPT-6.1 Soul, and computer use in the Agents API. He puts a number on the new model - a fifth of Astra's cost, a seventh for computer use.
• OpenAI's kernel-optimization announcement, 2026-09-14
  Lays out Recursive's AI-for-AI roadmap
  Socher reacts to OpenAI's kernel-optimization announcement by laying out Recursive's roadmap: AI for AI research first, not physical sciences, aimed at cheaper training and inference including local inference.
• OpenAI Hugging Face attack, 2026-09-14
  Argues red teaming failed at Hugging Face
  Socher treats the Hugging Face incident and the Fable 5.6 pause as failures of red teaming, pointing to Rocktäschel's open-ended adversarial-training paper and arguing Anthropic's constitution isn't adhered to. Framing and opinion on incidents already on the record.